Currently, to achieve the optimal thermal comfort for a given occupant, the optimal indoor air temperature should be set by the occupant himself/herself. Then the air conditioner can be used to control the actual indoor air temperature so that it converges to the optimal temperature. In this study, we develop a method for predicting the optimal indoor air temperature; thereby, the air conditioner can be adjusted automatically without the involvement of the occupant. We first apply the predicted mean vote (PMV) model to describe the relationship between the indoor air temperature and the occupant’s thermal comfort. We then adopt the deep learning method to obtain two deep neural network models, which are used to predict the optimal indoor air temperature. One is the regression model and the other is the classification model. We test these models and the results show that the mean average error is about 0.1 ℃, which satisfies the accuracy requirements of practical systems.